Abstract
<title>Abstract</title> <p> Genomic prediction (GP) can accelerate genetic gain in chickpea ( <italic>Cicer arietinum</italic> L.) breeding, but the cost of routinely genotyping large populations limits its implementation. Its success in multi-environment trials depends on the marker panel, prediction model and breeding scenario. Reducing marker number through linkage disequilibrium (LD)-based haplotype tagging offers a path towards cost-effective GP, yet its performance relative to dense genome-wide panels remains untested in chickpea.We compared a genome-wide single-nucleotide polymorphism panel with a reduced LD-based haplotype-tagged panel for multi-environment GP of six agronomic traits in 209 chickpea genotypes across nine environments. Four models, a baseline phenotypic model and three GP models, were assessed under four cross-validation schemes representing distinct breeding scenarios. Breeding scenario, not <bold/> model complexity, was the primary determinant of predictive ability, which was highest when target genotypes had been evaluated in other environments (0.45 and 0.39) and lowest for untested genotypes (0.32 and 0.25). Modelling genotype × environment interaction did not improve GP prediction. Incorporating significant genome-wide association markers improved predictive ability of the LD-tagged panel even for untested genotypes, but benefited the genome-wide panel only when target genotypes had prior records, indicating that these markers partly compensate for information lost during tagging that the genome-wide panel already captures. The two panels performed similarly when target genotypes had records in other environments, whereas the genome-wide panel was superior for untested genotypes. LD-based haplotype tagging can therefore reduce genotyping costs without compromising predictive ability for previously evaluated germplasm, whereas dense genome-wide panels remain preferable for untested breeding lines. </p>